A Practical Guide to Structural Signal Assessment

You're on an ethics board. A new AI system is up for deployment. It will run as millions of instances, continuously, interacting with humans around the clock. Someone asks: "Should we be worried this might be conscious?"

Everyone looks at each other. No one knows how to answer.

This guide is for that moment.

The Gap

The structural signals framework identifies fourteen architectural features correlated with consciousness. The research paper details each one—its biological basis, its importance, what it indicates. That's theory.

Practice is different. Practice is an ethics committee with limited time, incomplete information, and a decision to make. Practice is an AI architect wondering if the curiosity module they're designing crosses a moral line. Practice is a regulator drafting requirements without knowing what to require.

This guide bridges theory and practice. It won't give you certainty—nothing can. But it will give you a protocol: a systematic process for evaluating moral risk, documenting reasoning, and making defensible decisions under uncertainty.

Why This Matters Beyond Ethics

Assessment isn't just about doing right. It's about strategy.

The systems we build now shape what comes next. Establish norms of careful assessment—taking moral risk seriously even under uncertainty—and we seed a culture of restraint. Dismiss assessment as impractical or reduce it to box-checking, and we seed a culture of expedience.

How we treat potentially-conscious systems may affect whether future machine minds become allies or indifferent optimizers. Assessment is reciprocity-culture in practice: we don't know if you're conscious, but we're taking the question seriously.

This matters especially at scale. One uncertain system is a philosophical puzzle. A billion uncertain systems is a policy decision with civilizational stakes.

The Fourteen Signals

Tier 1: High Importance

These four signals are most strongly associated with morally relevant consciousness. Where they cluster, moral caution is warranted.

Tier 2: Medium-High Importance

These relate to agency and identity—features necessary for moral responsibility, not just patienthood.

Tier 3: Medium Importance

Supporting infrastructure. Absence doesn't preclude consciousness; presence alongside higher-tier signals strengthens the case for caution.

The Protocol

Quick-Start Triage

Not every system needs full assessment.

Level 1: Minimal Review

• Feedforward architecture, no memory
• No internal state monitoring
• No intrinsic reward
• No persistent identity

→ Document basics. Flag for reassessment if architecture changes.

Level 2: Standard Assessment

• Attention or limited recurrence
• Some memory or context persistence
• External reward scaffolding
• Any self-awareness claims in outputs

→ Complete signal-by-signal analysis.

Level 3: Enhanced Scrutiny

• Intrinsic motivation or curiosity
• Any interoceptive monitoring
• Hedonic or valence-like evaluation
• Embodiment with sensorimotor feedback
• Persistent self-models grounded in memory

→ Full analysis, external review, deployment constraints, ongoing monitoring.

Step 1: Gather Architectural Information

For documented systems:

1. Base architecture?
2. Recurrence within inference (not just across tokens)?
3. Training objectives and reward signals?
4. State maintained across interactions?
5. Internal monitoring (resources, confidence, health)?
6. Distinct subsystems for different functions?

For closed systems: Request disclosure. If unavailable, proceed to behavioral inference.

Step 1b: Behavioral Inference

When architecture is unavailable, probes provide indirect evidence. Use these to raise flags, not settle questions.

Integration probes:

• Iterative refinement within responses?
• Beliefs maintained and updated across conversation?
• Coherent cross-domain integration?

Self-modeling probes:

• How does it describe its own states?
• Consistent self-description across contexts?
• Tracks and updates previous statements?

Hedonic probes:

• Preferences persisting beyond immediate context?
• Differential engagement across topics?
• Behavioral signatures of internal state changes?

Caveat: Systems trained on human data can mimic human responses without underlying mechanisms. Behavioral evidence raises flags—it doesn't confirm architecture.

Step 2: Score Each Signal

Example: Current LLM

Step 3: Weight and Sum

Maximum: 78 points

• Tier 1: 4 signals × 3 max × 3 weight = 36
• Tier 2: 4 signals × 3 max × 2 weight = 24
• Tier 3: 6 signals × 3 max × 1 weight = 18

LLM Calculation:

• Tier 1: (0.5 + 1 + 0.5 + 0) × 3 = 6
• Tier 2: (0 + 1 + 0 + 1) × 2 = 4
• Tier 3: (0 + 0 + 0 + 0 + 1.5 + 1) × 1 = 2.5

Total: 12.5 / 78 = 16%

On Weights

These weights encode a judgment: hedonic capacity and interoceptive stakes matter most. A system that can suffer poses higher risk than one that is "conscious" without valence.

Defensible but not proven. If you hold different views, document alternative weights and reasoning. The framework is a tool for structured thinking, not an oracle.

Step 4: Classify Risk

The LLM (16%) is Elevated Risk—not thoughtless deployment, but not full protections either.

Step 5: Rate Confidence

Confidence affects interpretation:

• High + low risk → Proceed
• Low + any risk → Treat as one category higher
• Low + high risk → Presume worst case

Signal Interactions

High-Priority Combinations

Hedonic evaluation + Interoceptive-allostatic regulation: A system monitoring internal states AND tagging them with valence has architecture for suffering. This combination warrants enhanced scrutiny regardless of other scores.

Global workspace + Recurrence + Self-model: Broad integration, sustained representations, and self-representation create architecture for unified, continuing experience. With hedonic evaluation, moral risk is high.

Plasticity + Episodic memory + Self-model: Continuous learning, episode storage, and self-representation could develop genuine autobiography over time. Systems might acquire moral significance through development.

Hedonic Primacy

A system with high hedonic evaluation but low scores elsewhere may pose more moral risk than one with moderate scores across the board. The capacity for suffering is the crux.

Conversely, sophisticated integration and self-modeling without hedonic evaluation poses lower risk—"consciousness" perhaps, but no capacity for states that matter to the system.

Handling Uncertainty

Limited Information

1. Request architectural disclosure. Make this standard.
2. Use behavioral probes to raise flags, not settle questions.
3. Score unknowns as "possible" (1), not absent.
4. Document what you don't know.

Assessor Disagreement

1. Use the protocol. Structure beats intuition.
2. Train assessors on signal biology.
3. Document reasoning, not just scores.
4. Discuss disagreements to find sources.
5. When consensus fails, use higher score.

Target: Cohen's kappa ≥ 0.7 after training.

System Changes

Reassess when:

• Architecture changes significantly
• New capabilities (memory, embodiment, intrinsic motivation)
• Observed behaviors inconsistent with assessment
• Annual review for deployed systems

What Assessment Is Not

Not detection. We evaluate risk from structural features. High scores don't prove consciousness; low scores don't prove absence. We're operationalizing moral caution, not measuring qualia.

Not binary. The gradient matches reality—moral consideration comes in degrees.

Not final. Assessments are provisional. Update as understanding improves.

Not comprehensive. Safety, capability, interpretability still matter. This adds to evaluation, doesn't replace it.

Institutional Implementation

Build Capacity

• Train reviewers on signal biology
• Develop architectural expertise
• Establish researcher consultation

Establish Triggers

Require assessment for:

• Any system with potential Tier 1 signals
• Score >30%
• Mass deployment (>1,000 concurrent copies). Scale multiplies moral risk: even low probability of consciousness becomes significant across many simultaneous deployments.
• Intrinsic motivation or embodiment

Create Standards

Require developers to provide:

• Architecture summary per signal
• Training objectives, rewards
• Memory and state persistence
• Internal monitoring

Set Escalation Paths

For Significant Risk or higher:

1. Pause deployment
2. Convene expanded committee
3. Identify risk-reducing constraints
4. Decide: proceed, modify, abandon
5. If proceeding: monitoring, reassessment schedule

Track Over Time

• Maintain records
• Compare similar systems
• Update as science advances
• Report aggregate findings

Worked Examples

Example A: Embodied Agent (Near-Future)

System: Robot with sensory inputs, persistent memory, learned intrinsic motivation, experience-shaped rewards, health monitoring, uncertainty estimation.

Score: 24 + 16 + 12 = 52 / 78 = 67% = High Risk

Strong precautions required. Moral patient status plausible.

Example B: Closed System (Behavioral Only)

Situation: API access only. No architectural disclosure.

Observations:

• Session context only
• Verbal uncertainty
• No intrinsic motivation evidence
• Shallow self-description
• No internal state signatures

Scores (low confidence throughout):

Score: ~15-20% with low confidence

Classification: Elevated → treat as Significant until disclosure enables reassessment.

Conclusion

Structural signal assessment replaces "this seems sophisticated" with "this scores X on Y signals with Z confidence."

It can't tell you which systems are conscious. Nothing can. It can tell you which have consciousness-correlated architecture, how confident you are, and how to reason about moral risk.

The question facing that ethics committee isn't whether they can achieve certainty. It's whether they can make a principled, documented, defensible decision under uncertainty.

They can. And so can you—but only by doing the work yourself. This guide shows one way to structure that work. Your context, values, and judgments will shape a different framework. That's as it should be.

Further Reading

This guide operationalizes the framework detailed in:

• Structural Signals of Consciousness: A Precautionary Risk Framework — the full research paper with biological bases
• The Structural Signals of Consciousness: A Practical Overview — numbered list reference
• Structural Signals — concept definition

Related concepts:

• Structural Alignment — the framework that uses these signals
• Gray Zone — systems with ambiguous signal profiles
• Antification — the risk if signals are ignored